Truace
Subscribe
60/100

Moderate evidence

Cite
More
Export CSVExport JSONEmailLinkedInX

Rainfall forecasting AI still trails the official day-one outlook after 4 years of live tests

Study title: Progress toward forecasting excessive rainfall with random forests based on a deterministic convection-allowing model

Evidence type
Observational study
Sources
1 sourcePeer-reviewed
Published
2 min read
Truvace record
TRV-2026-2655cite or check it

Study title: Progress toward forecasting excessive rainfall with random forests based on a deterministic convection-allowing model

Researchers tested how to design a random forest system that predicts excessive rainfall from the High-Resolution Rapid Refresh (HRRR) weather model. Using data from all nearby gridpoints and averaging predictors across several model runs significantly improved forecasts, particularly for events with small-scale rainfall peaks.

Read the study abstract

Abstract Machine learning applied to high-resolution numerical weather prediction forecasts is gaining popularity for prediction of high-impact weather events, but questions remain on how to design such systems. Building off prior work, we describe sensitivity experiments using a random forest (RF) system for excessive rainfall prediction using the deterministic High-Resolution Rapid Refresh (HRRR). We find that applying an RF trained on HRRRv3 to daily forecasts from HRRRv4 does not significantly degrade result…

55Gain side

How the two sides scoreClose call

57Problem side

The gain

In sensitivity experiments on a random forest excessive rainfall system built on HRRR forecasts, using all nearby gridpoints and averaging predictors across several model initializations significantly improved forecasts, especially for small-scale rainfall maxima.

The problem

In more than four years of real-time forecasts, random forests based on HRRR rainfall forecasts underperformed a Global Ensemble Forecast System-based random forest, and neither matched the operational day-one excessive rainfall outlook; the HRRR version also underpredicted events.

A fixed formula sets both scores from the evidence and its sources. Reader votes don’t move them.

What the scores mean

G is the gain side, out of 100. It goes up when the gain is specific, when it measurably helped people and when the sources behind it are strong.

P is the problem side, out of 100. It goes up when the problem is specific, when it measurably hurt people and when the sources behind it are strong.

In brief

The system was evaluated during the Flash Flood and Intense Rainfall (FFaIR) Experiment over three consecutive years and showed year-over-year improvement. Over more than four years of real-time forecasts, however, it still underperformed a random forest built on the Global Ensemble Forecast System, and neither matched the operational day-one excessive rainfall outlook.

The finding shows that machine-learning post-processing of single high-resolution models has not yet overtaken existing forecast products for this task. The HRRR-based system also tended to underpredict excessive rainfall, partly because verification used a more comprehensive event dataset than training did.

Main points

  1. A random forest trained on HRRRv3 and applied to daily HRRRv4 forecasts did not significantly lose accuracy.
  2. Forecasts improved significantly when the system used all nearby gridpoints instead of sparse input points.
  3. Averaging predictors across multiple HRRR initializations improved forecasts because the random forest could learn from lower-uncertainty inputs.
  4. The system was evaluated in the FFaIR Experiment over three consecutive years, with year-over-year improvement.
  5. The HRRR-based random forest underpredicted excessive rainfall, partly because verification used a broader event dataset than training.
  6. Over more than four years of real-time forecasts, it trailed a Global Ensemble Forecast System-based random forest and the operational day-one outlook.

The full story

The work builds on earlier research using machine learning with high-resolution numerical weather prediction to forecast high-impact weather. The team ran sensitivity experiments on how to design the system, using the deterministic HRRR model as the source of predictors.

Two design choices mattered most: spatial aggregation of predictors across nearby gridpoints helped most for events with small-scale precipitation maxima, and time-lagged ensemble mean predictors from multiple HRRR initializations reduced uncertainty the random forest had to learn around.

Real-time evaluation covered more than four years and included the FFaIR Experiment across three consecutive years, where the system improved each year. Even so, it remained less skillful than a Global Ensemble Forecast System random forest and the operational day-one excessive rainfall outlook.

What this doesn’t fix

The HRRR-based system showed an underprediction bias, partly because verification used a more comprehensive excessive rainfall dataset than the one used for training.

Sources

1 on file
  1. Artificial Intelligence for the Earth Systemsdoi:10.1175/aies-d-26-0005.1
    Peer-reviewed2026-10-26

Talk it over

Four questions to start onThis record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

Start with a question

Rainfall forecasting AI still trails the official day-one outlook after 4 years of live testshttps://truvace.com/item/progress-toward-forecasting-excessive-rainfall-with-random-forests-based-on-a-de

  1. This record found a gain and a problem in the same place. Which matters more to you, and who might answer differently?

    The gain In sensitivity experiments on a random forest excessive rainfall system built on HRRR forecasts, using all nearby gridpoints and averaging predictors across several model initializations significantly improved forecasts, especially for small-scale rainfall maxima.

    The problem In more than four years of real-time forecasts, random forests based on HRRR rainfall forecasts underperformed a Global Ensemble Forecast System-based random forest, and neither matched the operational day-one excessive rainfall outlook; the HRRR version also underpredicted events.

    Answer in the thread

  2. How do we know? What can this kind of evidence show, and what can’t it?

    The evidence This rests on an observational study, from Artificial Intelligence for the Earth Systems.

    Answer in the thread

  3. What does it leave open? What else would you want to know before acting on it?

    The record’s own caveat The HRRR-based system showed an underprediction bias, partly because verification used a more comprehensive excessive rainfall dataset than the one used for training.

    Answer in the thread

  4. Have you seen AI used this way where you study, work or live? Did it go the same way?

    Answer in the thread

The questions are the same on every record. The details under them are taken from this record as written.

What readers say

More from Climate